acceptodds
Withdrawn from ICLR 2027

RAP: Phased Retrieval-Augmented Reasoning for Incident-Aware Traffic Prediction

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Traffic prediction plays a pivotal role in route planning and traffic management. Most existing methods focus primarily on historical spatio-temporal dependencies, overlooking external disruptions, particularly non-recurrent traffic incidents that can substantially alter traffic patterns. However, the unpredictability and heterogeneous impacts of incidents make it difficult to extract transferable incident-related knowledge from historical data, limiting the ability to accurately capture incident-affected traffic dynamics. To address these challenges, we propose RAP, a phased retrieval-augmented reasoning framework for incident-aware traffic prediction. RAP explicitly models incident-induced dynamics by integrating retrieval-augmented mechanisms with Large Language Model (LLM)-based reasoning. Specifically, it comprises two core components: Historical Knowledge Discovery and Context-Aware Phased Reasoning. The former extracts three complementary forms of transferable evidence from historical traffic data: incident patterns, normal counterfactual sequences, and experience trajectories. The latter uses an LLM to assess and consolidate heterogeneous evidence into a compact context, then partitions the prediction horizon into distinct impact phases for phase-consistent prediction. Extensive experiments on three real-world datasets demonstrate that RAP consistently outperforms state-of-the-art baselines, showing up to 5.27% performance improvement in incident-affected scenarios. The code is available at: https://anonymous.4open.science/r/RAP-608C.

Primary area:
learning on time series and dynamical systems
Keywords:
Traffic Prediction, Retrieval-Augmented Reasoning, Incidents

Reviews

Among:
Shape:
1234this paper, Bottom 25%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.75 over 4 reviews, above 25% of the others.
Reviewer_1oy5OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes RAP, a retrieval-augmented framework that combines historical incident knowledge, experience accumulation, and LLM-based phased reasoning to improve traffic flow prediction under non-recurrent traffic incidents.

Critical strengths

The paper has a clear motivation and presents a coherent overall story. The experimental results demonstrate some improvements over existing methods.

Critical weaknesses

W1. Limited evidence supporting the effectiveness and necessity of the core modules.

I think the effectiveness of the core method may need further experimental validation. In the ablation study, although the paper emphasizes the importance of Phased Prediction and Experience Accumulation, removing these two modules actually leads to a slight improvement in MAPE. While the performance in terms of MAE and RMSE decreases slightly, the overall differences are relatively small. Therefore, I think the authors should further clarify under what conditions these two modules are necessary and whether their improvements are sufficiently consistent. Since these modules are also among the main contributions of the paper, the current ablation results raise some questions about their necessity.

W2. Relatively high inference latency.

Although the authors provide some efficiency analysis in the appendix and reduce inference time through layering, the average processing time is still around 129 seconds per incident. I think this latency is relatively high for downstream applications such as traffic flow prediction, which can be time-sensitive. It would be helpful to further discuss the trade-off between accuracy and efficiency, particularly whether the accuracy improvements achieved by RAP are sufficient to justify the additional inference latency and computational cost compared with existing lightweight methods.

W3. Lack of analysis of experimental stability.

I did not find results reporting the variance or confidence intervals across multiple runs. Considering that the method involves LLM-based reasoning and that some of the reported improvements are relatively small, I think it would be useful to further evaluate the stability of the results.

W4. Potentially missing retrieval-augmented baselines.

Although the current baselines cover a number of traditional traffic forecasting methods, I think the evaluation could benefit from a stronger non-LLM retrieval-augmented baseline. For example, using the same historical retrieval information, a simpler retrieval-based correction or lightweight fusion approach could be considered. Such a comparison would help clarify how much of RAP's improvement comes from retrieval itself and how much comes from LLM-based reasoning. I think this distinction is important for demonstrating the necessity of the proposed method.

Questions for authors

N/A

Reviewer_NjHuOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

RAP is a retrieval-augmented framework for incident-aware traffic prediction that combines a base traffic forecaster with LLM-based reasoning. Given traffic observations and explicit incident information, it constructs three sources of evidence: historical incident cases, normal reference sequences derived from incident-free windows, and experience trajectories from completed prediction episodes. The framework organizes affected sensors into layers based on their road-network relationship to the incident. An LLM integrates the available evidence and divides the prediction horizon into phases to generate traffic forecasts. Experiments evaluate RAP on three regional subsets of TraffiDent, with comparisons across forecasting baselines, backbone models, and reasoning LLMs.

Critical strengths

S1. Understanding and predicting how traffic evolves following non-recurrent incidents is important for traffic management. The paper addresses this practically relevant problem by focusing on incident-induced disruptions and subsequent recovery dynamics.

S2. The paper compares multiple LLMs, providing insight into how the choice of reasoning model affects performance.

S3. The authors provide an anonymous code repository to support reproducibility.

S4. The paper is generally well organized, with a clear framework overview and detailed methodological descriptions.

Critical weaknesses

W1. RAP relies on explicit incident reports being available at prediction time. In practice, timely and complete reports may be difficult to obtain, limiting the framework’s applicability in settings with delayed, incomplete, or unavailable incident information.

W2. Prior research has addressed non-recurrent traffic dynamics even in settings without explicit incident annotations, including decoupling-based approaches that separate recurring patterns from irregular variations. Characterizing earlier methods as overlooking external disruptions is unreasonable. The paper should acknowledge these approaches and clarify how RAP advances beyond implicit modeling of aperiodic dynamics through explicit incident-conditioned retrieval and reasoning. Comparisons with representative decoupling-based methods would help establish whether the gains arise from access to additional incident information, the proposed reasoning mechanisms, or both.

W3. Section 5.1 lists PatchSTG and ASTGCN as the attention-based baselines, but does not explain the omission of representative alternatives such as GMAN, PDFormer, and STAEformer. PDFormer is particularly relevant because it explicitly models dynamic spatial dependencies and propagation delays, which are closely related to RAP’s motivation concerning the propagation and temporal evolution of incident impacts. STAEformer would provide another competitive comparison for assessing whether the reported gains persist against alternative spatio-temporal modeling approaches. Although the inclusion of PatchSTG is valuable, the current comparisons do not sufficiently establish RAP’s broader superiority. The authors should justify the baseline selection and include targeted comparisons, particularly with PDFormer and STAEformer, under the same data splits, forecasting horizons, and incident-evaluation protocol.

Questions for authors

Q1. The current experiments do not clearly separate the benefits of contextual retrieval from those of LLM-based reasoning. Could a non-LLM fusion or residual-correction model achieve comparable performance when provided with the same base forecast, incident information, and retrieved evidence? Such a comparison would help establish the additional benefit of RAP’s phased reasoning mechanism.

Q2. Why are RAP’s gains less consistent at 15- and 30-minute horizons than at longer horizons, despite access to explicit incident information? The reported 128.84-second runtime alone seems insufficient to explain this pattern. Please clarify the timing of incident-report availability relative to prediction and which components drive the gains across horizons.

Reviewer_oAaKOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

The manuscript proposes RAP, which first uses a traffic model to make predictions, then provides them to an LLM along with current observations, incident information, and information from past cases. The LLM divides the prediction horizon into phases and generates traffic flow predictions. Results on three regions show lower average errors than the evaluated baselines.

Critical strengths
  1. Traffic incidents can disrupt usual traffic patterns. Combining incident information with historical traffic observations is therefore a worthwhile prediction problem.

  2. The experiments compare several traffic prediction baselines and evaluate different traffic models and LLMs. The results across three regions suggest that the framework can improve prediction accuracy.

Critical weaknesses
  1. Much of the method consists of selecting and organizing inputs for an existing LLM, including traffic histories, foundation model predictions, and historical incident patterns in Eq. (2). The paper does not convincingly establish what methodological advance this provides beyond integrating these components for the traffic task. The key step from these inputs to numerical predictions is left to the LLM. Although the paper includes ablations and case studies, it remains unclear under what conditions these LLM-based revisions help or hurt prediction accuracy.

  2. Although different LLMs are compared, the experiments do not report variation across repeated runs or sensitivity to changes in the prompts. This makes it difficult to understand whether the reported gains are stable or depend on the particular LLM configuration used.

Questions for authors
  1. Could the authors use existing results or a simple controlled example to show when LLM-based revisions improve forecasts and when they make them worse? How sensitive are the results to prompts and generation settings?

  2. Could the authors clarify exactly what data RAP and each of the 11 baselines use for training and prediction? Do they use the same training data including incident information and historical observations? If the information differs across RAP and other methods, how is this accounted for in the comparison?

Reviewer_ZWvVOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

RAP starts with a frozen traffic forecaster and asks an LLM to refine its predictions using incident descriptions, recent traffic, historical incidents, normal references, and past prediction experience. The normal reference is a weighted future from incident-free windows at the same node. Historical incidents are retrieved by incident attributes, traffic sequences, and the joint state of the forecaster and the normal reference. After an outcome becomes available, the system can store the error and a reflection for future use.

Critical strengths

The best part of the paper is the evidence organization. Normal references, three views of historical incidents, and joint-state matching are combined in a concrete way. Eqs. (21)–(23) prioritize the worst mismatch across the forecaster, normal reference, and their gap, which is a sensible retrieval rule.

The paper also includes component ablations, five forecasters, several LLM backends, full prompts, and token measurements. The appendix is appropriately cautious: residuals are not causal estimates, and previous-layer predictions are only weak evidence.

Critical weaknesses

1. The remaining contribution is a specific organization of evidence, and its benefit has not been isolated

Several components have clear precedents. Jagadeesh et al., Application and Evaluation of Large Language Models for Forecasting the Impact of Traffic Incidents used an LLM for incident impact prediction. FUSE-Traffic combined event retrieval with a GNN. RAG4CTS used a healthy historical reference. MemCast and CastFlow already included experience memory, a forecast baseline, evidence-guided refinement, and reflection. RAP cites two of these papers but does not explain the difference in a way that establishes its own contribution.

The defensible increment is the particular combination of same-node normal references, three-way incident retrieval, and joint-state matching. I did not find that exact combination in the literature I checked. That makes it a useful candidate contribution, not evidence that it improves prediction. The key experiment should compare this organization with ordinary nearest-neighbor retrieval and simple correction rules under the same information and cost. CastFSR, CASTANET, and SEER are concurrent work and are not part of this deduction.

2. The phase and experience ablations do not isolate what is doing the work

With J=1 allowed, any 12-step vector can be placed in one phase. The schema itself therefore adds no new numerical expressiveness. The ordering of the prompt may still help, but w/o PP cannot separate the effect of adaptive boundaries from analysis text or generation budget. A direct comparison of flat output, fixed phases, and adaptive phases is needed.

The experience ablation has the same problem. Removing EA deletes historical outcomes, error information, and reflection text at once, so it cannot show that reflection helps. A cleaner comparison would keep the historical facts fixed and only add or remove the reflection. The ordinary sorted retrieval should also be compared with Eq. (23).

The normal reference is an empirical future from similar incident-free windows, not the potential outcome of the current incident had it not occurred. “Matched incident-free reference” would be a more accurate term than counterfactual. The paper should not claim that it separates a causal incident effect.

Questions for authors
  1. Does joint-state matching outperform ordinary nearest-neighbor retrieval when the candidate pool, information, and LLM budget are fixed?
  2. What happens if you compare flat output, fixed phases, and adaptive phases while keeping the analysis and generation budget constant?
  3. Does adding reflection text improve over the same historical facts without it? How do the normal reference, a weighted forecaster–normal blend, and historical residual correction compare?
  4. Can you report per-event paired uncertainty, preferably with block bootstrap by incident or date? MAE improves by 5.27% in San Francisco, while RMSE improves by only 0.33%. These metrics emphasize different aspects of performance, and neither aggregate number should be used alone to dismiss or establish the benefit.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.